18 Learning Action-Perception Cycles in Robotics fileLearning Action-Perception Cycles in Robotics 311 actions. These problems were supposed to be solved independently by general-purpose,

18

Learning Action-Perception Cycles in Robotics

A Question of Representations and Embodiment

Jeannette Bohg and Danica Kragic

Abstract

Since the 1950s, robotics research has sought to build a general-purpose agent capable of autonomous, open-ended interaction with realistic, unconstrained environments. Cognition is perceived to be at the core of this process, yet understan#ding has been challenged because cognition is referred to differently within and across research areas, and is not clearly defi ned. The classic robotics approach is decomposition into function-al modules which perform planning, reasoning, and problem solving or provide input to these mechanisms. Although advancements have been made and numerous success stories reported in specifi c niches, this systems-engineering approach has not succeeded in building such a cognitive agent.

The emergence of an action-oriented paradigm offers a new approach: action and perception are no longer separable into functional modules but must be considered in a complete loop. This chapter reviews work on different mechanisms for action-perception learning and discusses the role of embodiment in the design of the underly-ing representations and learning. It discusses the evaluation of agents and suggests the development of a new embodied Turing Test. Appropriate scenarios need to be devised in addition to current competitions, so that abilities can be tested over long time periods.

Introduction

In June 2014, the University of Reading reported that a machine passed the fa-mous Turing Test: a computer program impersonated a 13-year-old Ukrainian boy, called Eugene Goostman, and was able, through text interface, to make a

From “The Pragmatic Turn: Toward Action-Oriented Views in Cognitive Science,” Andreas K. Engel, Karl J. Friston, and Danica Kragic, eds. 2016. Strüngmann Forum Reports, vol. 18,

series ed. J. Lupp. Cambridge, MA: MIT Press. ISBN 978-0-262-03432-6.

310 J. Bohg and D. Kragic

suffi cient number of interrogators believe that they were communicating with an actual human being. This news attracted quite a bit of attention but not as much as one might have thought, since passing the Turing Test had been per-ceived to be proof that machines could think.

Turing proposed a general test of intelligence to measure the competency of an artifi cial system “in all purely intellectual fi elds.” He believed that by the year 2000, machines would be capable of this mental process, classically labeled cognition. He discussed the problems associated with deciding when a machine would convincingly reach this level and proposed that the ambiguous question of whether machines could think be replaced by an imitation game which the machine would have to win to prove cognitive competency (Turing 1950). This test, he imagined, would assess the intellectual capabilities of the agent independent of the actual mechanism or principle behind it. Turing’s original proposal and subsequent versions of the test (e.g., as is used in the Loebner Prize) did not attract signifi cant attention in the robotics community.

One possible explanation for this relative disinterest might be found in an interesting parallel (Russell and Norvig 2003) between the quest for artifi cial intelligence (AI) and artifi cial fl ight: Aeronautical engineering is not defi ned as making “machines that fl y so exactly like pigeons that they can fool even other pigeons.” Aeronautic researchers are interested in the principles of aero-dynamics. Thus, by analogy, AI researchers seem to have been interested in uncovering the underlying principles of intelligence rather than in duplicating an exemplar.

Early Approaches in Artifi cial Intelligence

In addition to suggesting the test, Turing (1950) theorized about the underlying principles. He favored the idea of a learning machine whose brain would be similar to that of a child (i.e., a blank slate). Certain built-in rules of operation for logical inference were possible, but these would be subject to change dur-ing learning. Interestingly, however, he did not consider it necessary for the agent to have limbs or eyes.

Subsequent researchers have dedicated signifi cant attention to the problems that a machine would face during the proposed imitation game: understand-ing and producing natural language text, representing general knowledge and information from the ongoing conversation, reasoning to answer questions or to draw novel conclusions, and learning from experience to adapt to new situ-ations. Special symbolic rule-based planners were developed that rely on the existence of an internal world model. Given a certain world state and a goal, these planners could devise a strategy to attain this goal. The fi rst planning system, STRIPS, was developed for the Shakey Robot at Stanford Research Institute (Fikes and Nilsson 1971) and functioned independently to how the robot built the world model, recognized certain objects, or executed planned



Learning Action-Perception Cycles in Robotics 311

actions. These problems were supposed to be solved independently by general-purpose, task-independent black-box modules. Signifi cant progress was made early on in terms of these planning algorithms. Based on this work, supercom-puters are now able to beat the best human players in chess or Jeopardy. Yet robots are still unable to demonstrate the autonomy and skill of a one-year-old child in terms of perception and motor control. We suggest that the greatest challenge to developing a general-purpose autonomous agent arises at the in-terface between the agent and the world, not at logical reasoning over readily given abstract symbols.

Emergence of the Action-Oriented Paradigm in Robotics

Moravec (1988) pointed out the following paradox: high-level symbolic rea-soning, which requires relatively high effort by humans, seems to be relative-ly easy to automatize. However, tasks that humans can perform effortlessly (e.g., grasping of arbitrary objects or manipulation of tools) seem diffi cult for machines to achieve. While we are consciously aware of symbolic reason-ing, these latter tasks are controlled by subconscious processes and thus they are much harder to reproduce. Moravec claims that these processes developed over thousands of years of evolution while abstract reasoning is a rather recent development.

Related to this, Brooks (1990, 1991b) proposed a new way to think about artifi cial intelligence. In contrast to Turing, he believed that a machine needs limbs and eyes to interact with a complex and dynamic environment. He reject-ed the focus on internal general-purpose representations of the world, symbolic reasoning, and a functional decomposition of intelligence. Instead, he defi ned intelligence in terms of a combination of simple behaviors, which were defi ned by directly connecting perception modules to controllers. These behaviors were combined in a structure that Brooks called the subsumption architecture. He could show that robots using this idea would expose intelligent behavior in dynamic and cluttered environments. They were even able to show simple grasps and navigate in mapped environments, without any need for complex internal representation and reasoning. These robots had no memory; instead they relied on sensors for continuous feedback from the world around them.

Parallels between Cognitive Science and Robotics

Almost simultaneously to Brooks’ proposal, the early work of Varela et al. (1992) established the “enactive approach” to cognition. Similarly to Brooks, Varela and colleagues did not consider cognition to be the process of extract-ing general-purpose, task-independent representations of the world. Instead, they held that cognitive processes of internalizing the external and building structures are guided by action. This is further related to the internal simulation




theory in which the brain simulates the environment and reasons on it before acting (Jeannerod 1988). Clark’s action-oriented representation (Clark 1998) and O’Regan and Noë’s (2001) sensorimotor contingency (SMC) theory both support this work. According to SMC theory, the agent’s SMCs are consti-tutive for cognitive processes and are defi ned as law-like relations between movements and associated changes in sensory inputs that are produced by the agent’s actions. Accordingly, “seeing” cannot be understood as the processing of an internal visual “representation”; seeing corresponds to being engaged in a visual exploratory activity, mediated by knowledge of SMCs. Additional evidence from psychology and neuroscience stipulates that action in biologi-cal systems participates as a generative model in perceptual processes and in structuring knowledge about the world (Gallese et al. 1996; Fadiga et al. 1999; Borroni et al. 2005).

Robotics Research Today

Although the proposal by Brooks is now widely considered to have marked a paradigm shift in robotics, it still remains to be shown whether these ideas can yield more high-level autonomous behavior than that of insects. However, the robotics community has placed more research effort on the interface between an agent and its environment. This does not mean that robotics has agreed on one approach. The classic approach to AI and new directions coexist and are potentially combined. As it happens, this situation is similar to the develop-ment that occurred in cognitive science (Engel et al. 2013).

Current research in robotics is largely shaped by the systems engineering approach (Brock 2011), which very often aims at solving problems related to a specifi c application. Within the current research funding landscape, prog-ress has to be fast and verifi able. Today’s robots are complex systems that require expertise in many different subjects. A roboticist may commonly be specialized in one of them and try to abstract away the others. For example, researchers who are experts in control may know little about visual perception. Therefore, these modules are abstracted away and treated as black boxes. Any potentially complex two-sided interaction between control and perception is replaced by a simplifi ed interface. Representations may be treated as general purpose and task independent. Very often, symbolic planners sit at the center of these approaches and devise plans that are computed over symbols provided by the black-box perception modules. Resulting action sequences are often ex-ecuted in an open-loop manner without checking to see whether the expected effect has actually been achieved.

The research area of computer vision is rooted in the demand of robot-ics research for general-purpose, task-independent representations of semantic entities (Horn 1986). Due to the diffi culties inherent in this problem, research in computer vision has developed away from robotics and now has little to do with it. First and foremost, it rarely addresses challenges that arise in robotics,




such as real-time requirements or the possibility to act in the environment for exploration. Similarly, in the area of control, the generation of movement is mainly studied in isolation. Feedback controllers usually close the loop around joint angles, velocities, or motor torques. Not as much focus has been placed on feedback about the environment structure. If this kind of feedback is re-quired, it is often provided by precise motion capture systems in the hope that sometime in the future computer vision researchers will deliver the promised reliable general-purpose black boxes.

These approaches have brought tremendous progress in their associated re-search areas. However, when trying to unite them within a robotics system through the systems- engineering approach, they are only successful in restrict-ed application scenarios that are not open-ended, largely static, and controlled. In general, current robots lag surprisingly far behind humans although they have faster and less noisy sensors and actuators and can perform rapid decision making and control (Wolpert et al. 2011). If we are still striving to discover the underlying fundamental principles between autonomous and purposeful behavior, we have not yet found the key.

Many people believe that the action-oriented paradigm offers the key to permit new insights. We have seen the emergence of the fi eld of developmen-tal robotics, which strives toward learning machines, already proposed by Turing. However, developmental roboticists also emphasize the importance of the learning agent being embodied (Lungarella et al. 2003). Below, we re-view a portion of the work that follows the action-oriented paradigm and focus mainly on mechanisms for action-perception learning. We focus on the role of embodiment in the design of the underlying representations as well as for the specifi c learning mechanism.

Representations

General-purpose autonomous robots cannot be preprogrammed for all the tasks they will be required to do; just like humans, they should be able to gather information from different sources and learn from the experiences of both humans and other robots. Thus, the ability to acquire new skills and adapt existing ones to novel tasks and contexts is a necessity. For some natural do-mains (e.g., cooking and meal preparation), models or plans for different tasks are already available. For example, web pages such as ehow.com and wiki-how.com provide simple and detailed instructions on how to plant a tree or make lemon curd. These sites contain thousands of directives for everyday activities: about 45,000 on wikihow.com and more than 250,000 on ehow.com. Using written and structured instructions is common for humans. Many repeti-tive and dangerous tasks in factories and laboratories have natural language and graphic workfl ow specifi cations that are similar to task instructions in the World Wide Web.




A natural idea is to enable robots to do the same. However, this poses sev-eral challenges. To look at, listen to, and perceive an instruction, robots need to be able to understand text, video, spoken commands, or even all of them at the same time. They need to be able to understand concepts that are symbolic and relate them to sensory information. For example, an object such as “fork” in spoken or written instructions needs to relate to specifi c visual features that can be extracted from an image. In addition, the representation of a fork needs to be such that the robot can distinguish it from a knife.

A classic approach is to develop general-purpose, task-independent rep-resentations of semantic entities needed in these aforementioned tasks. Representations like this promise effectiveness through compression of a lot of information into a single symbol, which then is able to generalize to all pos-sible situations and contexts.

Humans use visual and other sensory feedback extensively to plan and ex-ecute actions. However, this process is not a well-defi ned one-way stream: how we plan and execute actions depends on what we already know about (a) the environment in which we operate (context), (b) the action we are about to undertake (task), and (c) the result expected from our actions (effect). This insight has been picked up in robotics and resulted in many models that try to represent actions and percepts jointly instead of fi nding the one representation that matches all purposes. The concept of affordances, as proposed by Gibson (1977), has inspired representations, especially in grasping and interaction with objects.

To a certain degree, affordances can be observed in images. In several works (Bohg and Kragic 2009; Saxena et al. 2008; Stark et al. 2008), relations between visual cues and grasping affordances are learned from training data. In Stark et al. (2008), object grasping areas are extracted from short videos of humans interacting with the objects. In Bohg and Kragic (2009) as well as Saxena et al. (2008), a large set of two-dimensional object views are labeled with grasping points. Early work on functional object recognition (Rivlin et al. 1995; Stark and Bowyer 1996) can be seen as a fi rst step toward recognizing affordances from images. Objects are modeled in terms of their functional parts (e.g., handle, hammerhead; Rivlin et al. 1995) or by reasoning about shape in association to function (Stark and Bowyer 1996). In these approaches, the rela-tion between objects and action is usually predefi ned by humans.

As pointed out by Sloman (2001), we are not consciously aware of a signifi -cant amount of human visual processing: we do not experience using optical fl ow to control our posture nor are we aware of the saccades and fi xational eye movements that allow us to negotiate with the complexity of everyday scenes (Koch and Ullman 1985). Therefore, these processes are not easy for us to reproduce in an artifi cial system. It has been argued that representations should only be constructed by the system itself through interaction with and exploration of the world rather than through a priori specifi cation or pro-gramming (Granlund 1999). Thus, objects should be represented as invariant




combinations of percepts and responses, where the invariances (which are not restricted to geometric properties) need to be learned through interaction rather than specifi ed or programmed a priori (Granlund 1999). A system’s ability to interpret the external world is dependent on its ability to interact with it. This interaction structures the relationship between perception and action.

In robotics, this can be a slow process, due to the challenges involved when extensive physical interaction is required. Over the last several years, howev-er, advanced oculomotor and hand-eye systems have been demonstrated (e.g., Moren et al. 2008; Montesano et al. 2008; Kraft et al. 2008; Rasolzadeh et al. 2010). There are approaches that let the robot interact with its environment and learn through trial and error. One example is a cognitive model for grasp learning in infants (Oztop et al. 2005). A model for learning affordances using Bayesian networks embedded within a general developmental architecture has been proposed by Montesano et al. (2008). Kraft et al. (2008) proposed object-action complexes as semi-supervised procedures for encoding sensorimotor relations and showed how this can be used to improve the robot’s inner model and behavior. The idea was further developed by Song et al. (2010), where the relationships between object, action, constraint features, and task were en-coded using Bayesian networks.

These approaches often consider actions at discrete moments in time (e.g., a grasp when approaching but not yet making contact with the object) or as a discrete symbol (e.g., pushing, pulling, grasping, pouncing). The representa-tion of movement over time and how to couple it to sensory input is also an active area of research. One popular representation of this kind has been dy-namic movement primitives (Ijspeert et al. 2002; Schaal et al. 2007), proposed for both feedforward and feedback motor commands. Dynamic movement primitives relate to optimal control theory approaches such as minimum jerk trajectories (Flash and Hogan 1985), as well as machine learning approach-es such as hidden Markov models (Billard et al. 2004; Inamura et al. 2004). Dynamic movement primitives have been coupled to sensory input through maintaining a visual representation of the goal point and adapting the goal’s tracked position (Pastor et al. 2009). Other ways to shortcut perception have been to use motion capture systems or easy-to-detect fi ducial markers (Calinon 2009). Only recently have we seen how low-level sensory feedback can defi ne the goal directly (Pastor et al. 2011). During execution, a dynamic movement primitive is adapted such that the robot feels the same as when the movement was demonstrated.

Sensorimotor knowledge in humans is structured as of childhood, and this type of lifelong learning has been the focus of developmental approaches in robotics (Pfeifer and Scheier 1999; Kuniyoshi et al. 2003; Lungarella et al. 2003). Thus far, however, developmental approaches have been demonstrated on rather simplistic problems. If a large corpus of data is available, informed learning approaches, such as imitation learning (Schaal 1999), can be applied.




Learning and Priors

Much of the work on learning and priors was inspired by Piaget’s ideas of assimilation and accommodation. These two complementary processes of ad-aptation enable the experiences of the external world to be internalized. The problem of assimilation has been addressed more widely, given that some predefi ned structure has been used for classifi cation of new experiences. The problem of accommodation requires the representation of knowledge structure to be changed as the new data is gathered and requires more advanced learning techniques to be employed.

An organism cannot develop without some built-in ability. However, if all abilities are built in, the organism is unable to develop. There is an optimal level for how much phylogeny should provide versus how much needs to be acquired during the lifetime.

A human spends years interacting with its environment before it can mas-ter certain complex cognitive or motor tasks. At the same time, robots and the computational modules are often expected to learn from very little data. Imitation may be a very good way to bootstrap an artifi cial system. Even then, during its “lifetime,” a robot will encounter so many more situations than what could possibly have been demonstrated to it by a human teacher. In fi elds such as computer vision or speech processing, “ big data” (visual or auditory data annotated with strong or weak semantic labels) has become increasingly more available, impacting the very type of research that is being performed in these areas. Methods that can be trained on these massive amounts of data are cur-rently outperforming previous state-of-the-art approaches (Halevy et al. 2009). In robotics, there are no labeled databases of this order of magnitude to help bootstrap the system. This is most likely due to the complexity of a robot sys-tem, which makes it hard to collect and label these massive amounts of data. In contrast to computer vision and speech-processing data, a data point in a robotics database also depends on an action. Therefore, the usual assumption of independent and identically distributed data cannot be as easily made in a robotics system.

Robotic systems receive a continuous stream of sensory (visual, haptic, auditory) data that is currently largely unused. Data is often extracted at arbi-trary discrete points in time and then processed independently of the other data points in the time series. Exceptions to this are feedback controllers that enable robots to execute movements toward a goal or along a trajectory. Feedback on some state variable that is actively controlled is continuously gathered, and the appropriate action is computed to minimize the error between the actual and desired state. Most commonly, these state variables contain joint angles and velocities, forces, and torques which act on joints directly or on end-effectors. The state may also contain information from vision sensors, such as the pose of objects in the environment. For these quantities, good models (e.g., rigid body motion or dynamics) exist to help the controllers design and compute the next




best action based on carefully selected sensory feedback. For more complex or multimodal data or more complex goals, modeling the mapping between the sensory state and next best action becomes much harder.

Although the “big data” paradigm holds great promise for learning some of these aspects, it is unclear how the data which a robotics system produces (time-series, multimodal and synchronized in time, structured) can be lever-aged. Some examples exist for learning from time-series data, but the majority of work focuses on learning from discrete data samples.

Currently, the big data paradigm considers the problem of discovering cor-relations in data. However, robotics seems to be largely dominated by another structuring principle in data: causality. Discovering causality from data is dif-fi cult (Pearl 2009). Nonetheless, intuitively, understanding causality seems to be the key to predict the changes in sensory percepts after an agent executes an action.

Embodiment and Imitation

From the viewpoint of morphology, our bodies, actuators, and sensors exist to support effective action (Kuypers 1973) but there is nothing from the perspec-tive of robotic systems that requires a cognitive system to take human shape. Ziemke’s framework of embodied systems distinguishes fi ve types of embodi-ment: structural coupling, historical embodiment, physical embodiment, or-ganismoid embodiment, and organismic embodiment (Ziemke 2003). A single type of embodiment, however, cannot guarantee that the resultant cognitive behavior will be in any way consistent with human models or concepts.

Transfer of information between a teacher (human/robot) and a student (robot) requires a common knowledge representation. When the human and student have identical motor and sensory capabilities, the task may be simply to transform the action of one to the other by changing the frame of reference. Such transfers are not commonly possible, given that embodiments and asso-ciated capabilities often differ. To ensure compatibility with human concepts, there may be a need for higher similarity to humans regarding physical move-ment, interaction, exploration, and perhaps even human form (Brooks 2002).

In terms of object grasping and manipulation, the naïve approach to facili-tate grasp transfer between different embodiments is to model the observed ac-tion of the teacher and map all the action parameters to the robot hand, which is commonly referred to as the “action-level” imitation (Alissandrakis et al. 2002). However, since different embodiments have different capabilities, the action required to achieve the goal may be different.

Imitation learning is an effective approach for teaching robots simple tasks (Billard et al. 2008). The learning paradigm based on an internal model (Wolpert and Kawato 1998) has received considerable attention. The work by Rao et al. (2007) implements an internal model through Bayesian networks.




Demiris and Johnson (2003) show that the internal models that represent the brain circuitry subserving sensorimotor control also participate in action rec-ognition. They are used to predict the goal of observed behavior and activate the correct actions to maintain or achieve the “goal” state. Later work (Oztop et al. 2005) extends the use of an internal model to the domain of visual-manual tasks. We believe that future research in this area will address the interplay between the embodiment, knowledge representation, and learning in more de-tail. Abstraction from the embodiment may be a key, but one wonders to what extent this is reasonable to do.

Evaluation and Verifi cation: An Embodied Turing Test

While the aforementioned proposals have been verifi ed in specifi c scenarios and applications, we lack an understanding of how big their potential is toward the development of general-purpose cognitive agents capable of autonomous and open-ended interaction with realistic, unconstrained environments. How can an action-oriented approach actually be verifi ed and compared to other approaches?

Several tests have been proposed to evaluate general cognitive capabilities of an artifi cial agent, the most famous being the Turing Test. Turing (1950) was interested in the potential mechanisms and principles behind rational human reasoning, which we nowadays summarize with the somewhat fuzzy term of cognition. Instead of evaluating these mechanisms themselves, he proposed to measure the resemblance of an agent to a real person in a dialogue scenario. In this way, he proposed a way to circumvent the diffi cult problem of defi n-ing precisely the mental process of thinking. Turing believed that the exact computational structure of the mechanism does not matter as long as the ar-tifi cial agent is perceived to perform rational human reasoning. Furthermore, he believed that equipping the machine with a body was entirely beside the point. The actual Turing Test has played a signifi cant role in the fi eld of human-machine interaction. Although new versions of this test have been proposed (Harnad 1991; Marcus 2014), only the total Turing Test begins to test senso-rimotor capabilities.

To verify the methods proposed by the action-oriented paradigm, do we need a new embodied Turing Test?

One option would be to take the Turing Test and use it to evaluate not the actual mechanisms but rather the resemblance of how an artifi cial agent acts in the world compared to how a person would act. An external observer would decide whether a robot that is performing certain tasks in an environment is acting autonomously or is teleoperated. The tasks for the robot could involve manipulation and locomotion tasks of different degrees of diffi culty, in differ-ent environments (e.g., household or disaster relief scenarios). Tasks could also involve physical interaction or collaboration with other agents or humans (e.g.,




preparing a meal, clearing a dinner table, assembling furniture, rescuing a person from a disaster site, or collaborating with a person to perform assembly tasks).

The advantage of this type of test is that a specifi c task would need to be autonomously performed in a fl uid manner that resembles how a human would perform the task. We imagine this human manner to involve a certain level of dexterity, fl exibility in the presence of a dynamically changing environment, as well as robustness to noise and failures.

It is debatable whether the criterion of resemblance to human fl uidity when performing a task is desirable. It may be important in tasks where the robot is collaborating or interacting with humans such that its actions are predictable, but may have limited relevance when it comes to other (e.g., household) tasks.

A Robotics Challenge

Several robotics competitions have been set up to evaluate the performance of artifi cial agents, not only for purely intellectual tasks but also for tasks in-volving physical interaction: RoboCup, RoboCup@Home, DARPA Learning Locomotion, DARPA Autonomous Robotic Manipulation, and the DARPA Robotics Challenge. These challenges usually have well-defi ned goals that re-volve around a specifi c scenario, such as soccer, and can easily be verifi ed. The scenarios are usually formulated broadly so that the goals can be adapted and made more or less diffi cult from phase to phase. Furthermore, such competi-tions have the ability to bundle forces and focus them onto one goal (Marcus 2014). The spirit of competition seems to be a powerful source of motivation among researchers.

Do We Need a New Embodied Turing Test?

Without a doubt, it would be advantageous to have a test that could easily evaluate a set of well-defi ned goals. However, defi ning these goals poses the initial challenge. Although competitions can serve as a powerful motivator, experience shows that no matter how carefully such goals are defi ned or out of which original question they came, what counts in the end is winning. The hope or intention of discovering principles behind, for example, intelligence or autonomy, may be rejected in favor of getting the task done. Certainly this can be observed in earlier attempts to pass the Turing Test through the use of parlor tricks and purposeful deceit (Marcus 2014). All of the above-mentioned robot competitions encourage what is commonly referred to as hacking; that is, engi-neering solutions which exploit a fi xed structure in a scenario, thus sacrifi cing the generality of solutions. Nevertheless, competitions do offer clear demon-strations of which type of task can be achieved, and some fundamental insights are inevitably gained, although less than what one would hope for or expect.

Once a problem has been solved and its inner workings computationally and algorithmically revealed, we often no longer believe that the associated




artifi cial system is intelligent. It is just computation. We wonder whether it is appropriate to let a person judge the resemblance of an agent executing a task to a real human doing the same. This may shift the focus from fulfi lling a spe-cifi c task to that of doing it robustly and fl uidly.

Conclusions

Much can be said about perception and action as well as the work that has been done over the last sixty years. Here we reviewed cases which show some of the relations across different fi elds of research. Importantly, cognition is a process that needs to be studied and approached as such. Proper representa-tions and learning mechanisms are necessary to meet the goal of developing autonomous agents capable of open-ended interaction. Equally important is the issue of how to assess and verify that an agent has made the proper choices. To evaluate and verify the capabilities of a robot, we suggest that the well-known Turing Test be reworked into an embodied Turing Test. Many scenarios can be envisioned for such a test; however, we believe short, competition-like scenarios are insuffi cient. Appropriate scenarios need to be devised that will test a robot’s ability over long periods of time.



Documents

18 Learning Action-Perception Cycles in Robotics fileLearning Action-Perception Cycles in Robotics 311 actions. These problems were supposed to be solved independently by general-purpose,